SnowPro Specialty: Gen AI Question 144
Single answerUse fully-managed LLMs, RAG, and text-to-SQL servicesA retail company stores product manuals, return policies, and support playbooks in Snowflake and wants to build an internal assistant for customer support agents. The assistant must: (1) answer questions grounded in the latest approved documents, (2) minimize operational overhead by using managed capabilities where possible, and (3) allow agents to ask natural-language questions such as "Which products had the highest return rate in Q1 by region?" against curated warehouse tables. Which solution best meets these requirements?
- A
Use Cortex Search over the approved document corpus for retrieval-augmented answers, and use Cortex Analyst for natural-language-to-SQL questions over curated semantic models.
- B
Fine-tune a custom open-source LLM outside Snowflake on the document corpus, and generate SQL directly from prompts without defining any semantic layer.
- C
Use Cortex Complete alone for both document grounding and text-to-SQL, because a general LLM can infer the correct documents and SQL from raw prompts.
- D
Export documents to an external vector database and build a custom retrieval pipeline, then use ad hoc SQL generation against base tables to avoid creating governed models.
Show answer and explanation
Correct answer: A
Explanation
The best answer is to combine Snowflake's managed services according to the task type: Cortex Search for retrieval over trusted document content and Cortex Analyst for natural-language analytics over structured data. This mirrors recommended GenAI architecture patterns in Snowflake: use retrieval services to ground LLM responses on enterprise content, and use semantic-layer-driven text-to-SQL for reliable analytics questions. A general completion model alone is not enough for robust RAG or governed text-to-SQL. Likewise, custom external pipelines and direct prompt-to-SQL approaches may work technically but add unnecessary operational complexity and governance risk. In Snowflake documentation and product guidance, Cortex Search is positioned for AI-powered search and RAG experiences, while Cortex Analyst is positioned for business-user natural-language querying of structured data through semantic models.
- A. Correct.
Correct. Cortex Search is the managed Snowflake service designed for low-latency retrieval over enterprise content and is appropriate for grounding responses in approved documents as part of a RAG pattern. Cortex Analyst is the managed text-to-SQL service intended for natural-language questions over governed data, using semantic models to map business terms and metrics to the correct tables and logic. This combination satisfies all three requirements: grounded answers, low operational overhead, and natural-language analytics.
- B. Incorrect.
Incorrect. This introduces significant operational overhead and does not align with the requirement to prefer fully managed capabilities. In addition, generating SQL directly from prompts without a semantic layer is risky because business definitions, joins, and metric logic can be misinterpreted. Cortex Analyst is specifically intended to reduce those risks through semantic models.
- C. Incorrect.
Incorrect. Cortex Complete provides access to LLM inference, but by itself it does not replace a managed retrieval service for grounding on approved enterprise documents, nor does it provide the governed text-to-SQL behavior of Cortex Analyst. Relying on a general LLM alone is a common misconception; without retrieval and a semantic model, answers can be ungrounded and SQL can be unreliable.
- D. Incorrect.
Incorrect. Although a custom external retrieval stack could work, it does not meet the requirement to minimize operational overhead when Snowflake provides managed services for retrieval and text-to-SQL. Querying base tables with ad hoc SQL generation also bypasses governed business semantics, increasing the chance of incorrect answers and inconsistent metrics.